Wetland vegetation mapping improved by phenological leveraging of multitemporal nanosatellite images

Abstract

Accurate remote sensing of wetland vegetation is challenging due to heterogeneous land cover, dynamic water reflectance and extent, and spectrally similar plant types. The high spatiotemporal resolution of nanosatellite constellations enables ‘phenological leveraging’ (PL)—identification and tracking of spectrally distinct phenological events to help differentiate wetland vegetation. We use PlanetScope Dove-R images, field training data, and PL to map wetland landcover in a complex riverine wetland in Rhode Island, USA. Maximum Likelihood (MLC), Support Vector Machine (SVM), and Artificial Neural Network (ANN) classification algorithms are tested on individual, monthly- and multi-seasonal composite images. The greatest improvements in classification accuracy derive from targeting optimal months of maximum phenological contrast for each landcover class, and then compositing these optimized classifications into a single map. We conclude that nanosatellite constellations offer a powerful new approach for mapping wetland vegetation classes at fine spatial scales.